descriptinc/descript-audio-codec

What it solves

Descript Audio Codec (.dac) provides high-fidelity audio compression that significantly reduces file sizes without sacrificing sound quality. It addresses the problem of needing a universal codec that works across different audio domains—such as speech, music, and environmental sounds—while maintaining a very low bitrate.

How it works

The project uses an improved RVQGAN (Residual Vector Quantized Generative Adversarial Network) to compress 44.1 kHz audio into discrete codes. This allows for a compression factor of approximately 90x, achieving a target bitrate of 8 kbps. It can be used as a drop-in replacement for other codecs like EnCodec in audio language modeling applications (e.g., AudioLMs, MusicLMs).

Who it’s for

• AI Researchers and Developers: Those building generative audio models (like MusicGen or AudioLMs) who need a high-quality, efficient discrete representation of audio.
• Audio Engineers: Users looking for high-compression, high-fidelity neural audio codecs.

Highlights

• High Compression: Achieves ~90x compression of 44.1 kHz audio at 8 kbps.
• Universal Application: Works across speech, music, and environmental audio.
• Flexible Sampling Rates: Supports native weights for 16 kHz, 24 kHz, and 44.1 kHz.
• Integration Ready: Designed as a drop-in replacement for EnCodec for audio language modeling.